Predicting Transfer Values in the English Premier League
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Predicting Transfer Values in
the English Premier League
Dylan Newman1
Professor Peter Arcidiacono, Faculty Advisor
Professor Kent Kimbrough, Seminar Advisor
1
The author of this paper graduated from Duke University in 2016 with an economics and
history double major. He can be reached at dylanholtnewman@gmail.com.Newman 2
Table of Contents
Abstract 3
Acknowledgements 4
Introduction 5
Literature 10
Theoretical Framework 12
Data 20
Empirical Specification 21
Results 22
Overall Contribution and Further Effects 42Newman 3
Abstract
This paper examines factors that affect the transfer value of players transferred into the
English Premier League from 2009-2015. The analysis begins by examining what factors are
significant in determining a player’s projected transfer fee based on the website
Transfermarkt.com as well as the actual fee that the player was sold for. The paper goes on to
find that competition level and a player’s form are not statistically significant in models built to
determine a player’s transfer value. Quantile regression is then used to illustrate that there is a
superstar effect with a forward’s goal’s scored in the transfer market.
JEL Classification: Z21, L83
Keywords: Soccer, English Premier League, Transfer Fee, Quantile RegressionNewman 4
Acknowledgements
I would like to thank my seminar classmates for their invaluable feedback throughout the year. I
would also like to thank Aaron Libratore for his help with using Python to extract data, and Abby
Snyder for her help in data collection. This project would not have been possible without the
support of my family and the Duke Economics Department’s generous grant that funded my
research.Newman 5
I. INTRODUCTION
Sherwin Rosen (1981) argued that “small differences in talent become magnified in
larger earnings differences, with great magnification if the earnings-talent gradient increases
sharply near the top of the scale.” Given that true “superstars” are in short supply and in high-
demand, marginal increases in talent can lead to an extraordinary increase in value. Building off
this analysis, Moshe Adler (1985) believes that luck and positive social network externalities are
also a big proponent of stardom. Adler expanded on Rosen’s findings, stating that even among
those who have equal talent, one can be more of a star due to positive social network
externalities.
While in areas such as music, talent can be hard to identify due to there being a large
subjective component (Connolly and Krueger 2006), in sports, it is much more feasible to
quantify talent, given the amount of individual statistics collected. Better athletes tend to put up
better statistics and better teams tend to win games at a higher percentage. Players and their
performances tend to regress to the mean, causing luck to average out in the long run. While
studies supporting Rosen and Adler’s notions have been done in fields such as business, art,
music, and sporting leagues such as the National Hockey League, this extension of the superstar
effect has not made its way to English Premier League soccer, the top flight soccer division in
England and arguably the world. However, Franck and Nuesch (2006) examined the superstar
effect in the German Bundesliga, Germany’s top-flight soccer league, and found there to be a
strong superstar effect, where a small number of top players have transfer fees much higher than
other players. This paper will set out to examine if soccer players purchased by English Premier
League clubs with marginally better stats than others go for a much higher transfer fee.
Just this summer, English Premier League clubs combined spent over 700 million euro in
the transfer market. Clubs traditionally spend approximately 20% of total revenue per season inNewman 6
the transfer market; however, this can vary depending on team needs.2 Most notably, this past
transfer window, FC Barcelona sold Pedro for 30 million euros to Chelsea. Pedro is considered a
star player who has won a World Cup and European Cup with the Spanish National Team as
well as a Champions League Final with Barcelona. Meanwhile, nineteen-year-old Anthony
Martail of Monaco was sold to Manchester United for 80 million euros, despite only scoring
eleven goals for Monaco and having just one appearance for the French National Team. At first
glance, one may wonder why the player’s transfer values were as stated; however, advanced
statistical analysis used by the clubs in the transfer process likely explains the difference in the
two transfer fees.
Table 1: Transfer Window History
2
Swiss Ramble. Manchester United What Difference Does it Make? September 2015.Newman 7
Figure 1: Average English Premier League Transfer Fees Over Time
Average
Transfer
Fees
16
Transfer
Fee
(in
millions
of
euro)
14
12
10
8
Winter
6
Summer
4
2
0
2009
2010
2011
2012
2013
2014
2015
Year
In the global soccer market, teams can buy players from soccer clubs all over the world.
In England, this happens during a two-month window in the summer and a one-month window in
the winter. The summer window happens before the season starts, and runs from the first of July
to the first of September. The winter transfer window opens roughly midway through the season,
from 1 January to 1 February. Most major signings take place in the summer window, while the
midseason winter transfer window tends to be used as an opportunity to purchase reinforcements
incase a team realizes that they have a glaring need or have had injuries negatively affect the
team.3 In order to go about this transfer, clubs must pay the other club a fee for the value of a
player, known as a transfer fee, and the player must agree to the move. This transfer fee paid is
3
Transfer Window in European Soccer Explained. July 2014.Newman 8
different than a player salary. Each year, roughly 80 players are transferred to the Premier
League for a disclosed fee. Using this data, I can draw conclusions about the incremental effect
of statistics on a player’s market value (i.e., the incremental effect of a goal). Table 1 and Figure
1 provide descriptive statistics on the amount of players transferred to the Premier League each
season as well as the median and mean values of the transfer fees. From the summer of 2009 to
the summer of 2014, we see the average transfer fee rise year-to-year, presumably given the
growth of club’s revenue from commercial deals and broadcasting rights. Interestingly, the
average transfer fee slightly declines in 2015; however, due to a new Champions League
television deal, it is expected that the average transfer fee will increase in the coming windows.4
To analyze the effect of player statistics on transfer fees, I used the detailed statistics of
STATS LLC, a sports data collecting service that collects a wide variety of advanced statistics. I
examine the transfer fees of player purchased by English Premier League clubs, reported by the
BBC, since the end of the 2008 season. This coincides with the last major sale of an English
Premier League club, when Manchester City was sold to Sheik Mansour of Abu Dhabi. This
purchase is seen by many to have disrupted the transfer market, as Manchester City began
spending exuberantly and have the highest Premier League net spend on players since 2008.5
The 2009-2015 window range also includes 466 players that were transferred into the English
Premier League, including 441 outfield players.
Using these actual transfer fees as the dependent variable, and using the statistics from
STATS LLC as the independent variables, I examine the relationship between player statistics
and transfer values. I examine whether these to see if there is a superstar effect present in the
English Premier League’s transfers by using quantile regression to see the value of an
4
The Guardian. BT Sport Champions League Exclusive Rights. 9 November 2013.
5
Soccerlens. Manchester City Changing Face of English Football. August 2014.Newman 9
incremental goal for a player who is in the top 5% of goals scored compared to a player who
scores the median number of goals. In the 2014-2015 season, 977 goals were scored in the
English Premier League. As each team has a senior roster of 25 players, usually consisting of 22
outfield players, this means the average number of goals scored per Premier League outfield
player is 2.2. However, this is taking into consideration defenders and the members of the roster
who typically do not play very frequently, both who seldom score on average.
I focused solely on transfers to the English Premier League given the complexity of each
league’s labor laws and homegrown quotas, where a certain percentage of players on the team
must have citizenship of the country in which the league is located in. For example, the English
Premier League states that there must be four players on one’s 25 person roster who are English,
and four more who trained with an English club for three years before their 21st birthday,
regardless of nationality. The Italian Serie A states that there must be eight homegrown players,
four of which are Italian and four of which are Italian and trained with an Italian club for three
years before their 18th birthday, and is considering upping the quota number.6 Furthermore, the
Russian Premier League simply states that four players on each roster must be Russian, but is in
talks to dramatically increase this number before they host the 2018 FIFA World Cup.7
No rules have been changed regarding transfer fees or nationality quotas for the English
Premier League; however during my 2009-2015 window; however, television money for the
Champions League (in which the top four English teams participate in), as well as for the
English Premier League has massively increased over the past few years after new worldwide
contract negotiations. This leads many to believe that transfer fees are increasing at a rapid rate
6
Four Four Two. Premier League and Serie A: Radical Quota Changes Could Save Italian
Football.
7
New York Times. Russian League to Cut Back of Foreign Soccer Players. April 2015.Newman 10
year over year, as club’s revenues rise, which I will account for in my research.
For the rest of the paper, I go in detail about previous literature in section two, my
theoretical framework in section three, the data in section four, my empirical specification in
section five, my results in section six, and the overall contributions in section seven.
II. LITERATURE
In the early 2000s, the Oakland Athletics of Major League Baseball began to use a
“sabermetrics,” approach to better their team, as they did not have the monetary resources of
other teams. In this approach, they analyzed player’s statistics and found overvalued and
undervalued players on the market by finding statistics that were highly correlated with win
percentage. This is regarded as the first time that econometrics methods were used in sports.
While the Oakland Athletics never officially released their methods to the public, many
economists such as Hakes and Sauer (2006) believe that simple regressions can arrive at similar
findings. Since then, many sports teams have attempted to use advanced statistics to determine
optimal, undervalued players; however, many criticize this method, saying statistics cannot fully
explain sporting performance or that sports other than baseball are not optimal for this statistical
modeling. Baseball’s individualistic nature on offense, where only one batter is up at a time,
makes it easy to track, and therefore analyze, statistics for given players. Bill Gerrard, one of the
founders of the “moneyball method” that the Oakland Athletics implemented, and Howard
(2007) believe that the moneyball approach can be applied to other sports and leagues as well, as
long as the data is present and team executives and fans are not resistant to change.
This moneyball analysis has been extended to multiple sports across the globe, from hockey
to Australian Rules Football (Stewart, Mitchell, and Stravos, 2007); however the analysis seems
to be under-utilized with regards to soccer, with just one major paper having been written
regarding soccer. While most moneyball analyses revolve around baseball, Egon Franck andNewman 11
Stephan Nuesch (2012) used empirical tests to determine what individual player statistics were
most correlated with a team’s win in the German Bundesliga, the top professional soccer league
in Germany. Not surprisingly, they found goals and assists to be highly correlated with a win.
They also found shots on target and cross success rate to be significantly correlated with a win,
while red cards, yellow cards, and conceded penalties were highly correlated with a loss. Then,
they used these findings, which statistically affect the outcome of the game, as the explanatory
variables to better understand a player’s market value. Their analysis shows that the
aforementioned statistics and a player’s media popularity, measured by a number of times a
player was mentioned for non-performance reasons in popular German newspapers and
magazines, can explain roughly 70 percent of a player’s predicted transfer fee, as projected by
Transfermarkt.
Both Rosen and Alder argued that the value of superstars increases exponentially compared
to that of an average performer or artist. Alan Kruger (2005) examines and extends this superstar
effect when it comes to musicians – people will pay a much higher amount to listen to a superstar
artist than they would to listen to an average artist. This analysis is further extended to CEO pay,
where it is found that companies will pay more for a “superstar” CEO. Fortune 50 CEOs make
four times as much on average than an average Fortune 500 CEO.8 Extending this, in many
fields, the best are much more expensive than those who are just very good.
Using a quantile regression approach, as used by Franck and Nuesch, the incremental impact
of a goal can be analyzed for different stardom levels. For example, Franck and Nuesch find that
a star player at the 95% quantile sees his market value increase by 4.5% for each goal scored,
8
"How Superstars' Pay Stifles Everyone Else." New York Times. New York Times, 25 Dec.
2010. Web. 5 Oct. 2015.Newman 12
whereas an average player sees his value increase by approximately 1% for each goal scored. I
extend this analysis to the English Premier League, where there are arguably even more
superstars than the German Bundesliga.
I use a transfer fees as the dependent variable and have different player characteristics as the
independent variables, similar to what Franck and Nuesch’s analysis. While Franck and
Nuesch’s analysis will serve as a good template, I believe that there are potential additions and
improvements that will make my analysis stronger. These improvements, such as using an actual
transfer fee and examining the effect of competition level and form, will hopefully lead to having
a higher adjusted r-squared, thereby explaining more of the variation in the data and being a
better predictive model. In what follows, some potential improvements to the previous literature
are discussed.
III. THEORETICAL FRAMEWORK
a. Dependent Variable
Franck and Nuesch used projected market values made up by German newspaper Kicker
as the dependent variable in their model. While the German newspaper they used solely
predicted transfer values of all German players for that year, a leading soccer website also used
by Franck and Nuesch, known as Transfermarkt, estimates the value of soccer players from
around the world. These Transfermarkt estimates seem random at times, so I started by running
correlation tests of what they say the “market value” of a player is at the time that they are
transferred (what they predict the player should go for if they were transferred) with the
corresponding actual transfer value paid by the Premier League club to see how good these
estimated market values are. As the transfer fee rises, the correlation between the actual transfer
fee and the projected transfer fee diminishes. As such, using actual transfer fees as a dependent
variable should lead to better findings. Regressing the actual transfer fee with Transfermarkt’sNewman 13
projected value for a player at the time that they were transferred shows an adjusted R-Squared
of 73% and the slope coefficients are significantly different than one (Table 2). More so, when
doing a correlational analysis, we see that Transfermarkt’s projected values are 85.6% correlated
with the actual transfer values of a player (Figure 2).
Table 2: OLS Regression with the Transfer Fee as the dependent variable and the Transfermarkt
projection as the dependent variable
Transfer Fee Coefficient SE
(in tens of millions)
Transfermarket Projection 0.883*** 0.025
(in tens of millions)
Constant 7.948*** 0.308
N=441
Adjusted R-Squared=0.73
When looking solely at transfers that went for 10 million euro or higher, we see an even
poorer relationship between Transfermarkt’s projected values and the actual transfer fees paid. In
this case, the adjusted R-squared is 55% and the slope coefficients are significantly different than
one (Table 3). Running a correlational analysis between Transfermarkt’s projected transfer
values and actual transfer fees worth more than ten million euro in the past year saw that the
values are 75% correlated. As Transfermarkt becomes less accurate with more expensive transfer
fees, I believe the superstar effect and quantile regression will help me build a better model to
predict actual transfer values.
Although using Transfermarkt data would allow me to have a larger sample size, I will
still have an amble sample size and have a more reliable and useful model if the dependent
variable is the actual transfer fee. To extend Franck and Nuesch’s 2006 paper and for comparisonNewman 14
purposes, I run an initial regression using Transfermarkt’s “projected” transfer fees at the time
the player was transferred for players transferred into the English Premier League before running
regressions with actual transfer fees.
Table 3: OLS Regression with the Transfer Fee as the dependent variable and the Transfermarkt
projection as the dependent variable, only including transfers that had a transfer fee over 10
million euro
Transfer Fee Coefficient SE
(in tens of millions)
Transfermarket Projection 0.682*** 0.025
(in tens of millions)
Constant 1.253 0.308
N=133
Adjusted R-Squared=0.55
b. Player Characteristics and Statistics:
Player statistics are added to the model as independent variables as one would expect a
player’s transfer fee to be highly dependent on player’s statistical performances. These statistics
can be seen in Table 4.
While overall statistics are an important variable, breaking the statistics down by season
would portray a more accurate picture on how the player is doing. For example, someone who
scored 20 goals in his first season, two goals in his second season, and two goals in his third
season is different than someone who scored two goals in his first season, two goals in his
second season, and 20 goals in his third season, despite them both scoring the same number of
career goals. To best capture form, I try two dummy variables, one for assists and one for goals,Newman 15
equal to one if the output in the respective category increased from the previous year. I also try a
variable whose value is equal to the net change in goals and assists compared to the previous
year.
More so, a homegrown dummy variable is also included in the regressions. In Alex
Bryson, Giambattista Rossi, and Rob Simmons’ “The Migrant Wage Premium in professional
Football: A Superstar Effect?” (2007) it is shown that in the Italian Serie A, foreign players get
paid more than domestic Italian players, when controlling for players’ statistics. While
examining this for the English Premier League, I would hypothesize that the opposite could be
true, for transfer values, as eight players on a team’s 25-man roster must be from England or
have been in an English academy for three years before his twenty-first birthday. This quota
artificially raises the demand for English footballers, and as a result, raises the transfer fee.
Managers of Premier League clubs, such as Manuel Pellegrini of Manchester City, argue that
homegrown players are too expensive given the perceived inflated fee. In my regression, I would
add a variable for being an English “homegrown” player to see if, holding all else equal, being
homegrown leads to a higher transfer fee. Using an OLS model, I examine what exactly the
homegrown premium is, numerically speaking. Despite this “homegrown premium” being
heavily talked about in the media, the extent of it has not been analyzed before.Newman 16
Figure 2: Graph Depicting Transfermarkt’s Projected Transfer Values (in millions of Euro) on
the x-axis compared to the Actual Transfer Values (in millions of euro) on the y-axis for players
transferred to the English Premier League from the 2009 to 2015 season
50
40
ProjTransferFee
20 10
0 30
0 20 40 60
FeeNewman 17
Table 4: Player Characteristics and Statistics
Variable Description
Age Player’s age in years at the time they were
transferred
Age Squared The squared value of a player’s age at the time
they were transferred
Appearances The number of matches that a player has
appeared in
Matches Started The number of matches that a player has
started
Minutes Played The amount of total minutes that a player has
played in
Goals The number of goals scored
Assists The number of assists, passes that lead directly
to a goal, that a player has
Form Dummy Equal to one if a player has increased output in
goals or assists
Form Net Change Equal to the amount of goals and assists that
the player increased their output by compared
to the previous year
Homegrown Dummy Equal to one if the player is considered
homegrown in EnglandNewman 18
c. Purchasing Characteristics:
Purchasing characteristics are added to the model to control for the time that the player
was purchased and who purchased the player. As there are two transfer windows each year in the
English Premier League, a dummy variable will control for which time of year the player was
purchased. The winter transfer window is during the middle of the season, and many speculate
that this leads to higher prices as teams would want to charge a premium if they were to give
away a player without proper time to get a new player accumulated to the system.
Team fixed effects were also added to the regression, including all thirty-two teams that
have participated in the English Premier League since the 2009 season. The regressions include
fixed purchasing team effects due to the massive differences in revenue among teams. For
example, in 2014, Manchester United brought in 433 million euros, while Cardiff City brought in
just 83 million euros, by contrast.9
Another aspect I examine is if richer teams are charged a premium to buy players. It is
often brought up that more successful, richer teams always have to overpay for talent as the
selling teams know that they have a lot of money. The revenues of Arsenal, Chelsea, Manchester
City, and Manchester United outpace the rest of the teams in England, and are about double that
of the revenue of the sixth highest club, Tottenham Hotspur.10
9
Swiss Ramble. Manchester United What Difference Does it Make? September 2015.
http://swissramble.blogspot.com/.
10
Swiss Ramble. “Arsenal Searching for Hows and Whys.” September 2015.Newman 19
Table 5: Purchasing Characteristics
Variable Description
Purchasing Team Fixed Effect Dummy Dummy variables for the purchasing club
Variables
Selling Team Fixed Effect Dummy Variables Dummy variables for the selling club
Winter Transfer Window Dummy A dummy variable equal to one if the player
was purchased in the winter transfer window,
rather than the summer transfer window
Rich Club Premium Dummy Was the player bought by a team in the
Champions League
d. Popularity
Popularity, as shown by Franck and Nuesch, also plays an important role in the transfer
fee of a player. While I won’t have time to delve through the amount of non- performance related
articles that players are mentioned in, I tried to measure popularity based off of other online
metrics. Unfortunately, Facebook likes do not seem to be tracked over a long period of time and
Google searches are indexed to the peak of searches for that term. More so, while jersey sales
would be a good proxy, many leagues around the world do not list jersey sale data for individual
players. Similarly, autograph value could be a good proxy for player popularity, but it is difficult
finding autograph prices at the time a player was transferred. Further, Twitter mentions seemed
to be a potentially quantifiable metric; however, one can only go back three to five days for
Twitter data analytics, again making this a useless statistic for transfers that have previously
occurred. In short, there are many ways to measure popularity of players, but none seem to be
feasible at this time, which will be one downside of my analysis compared to previous analyses.Newman 20
IV. DATA
For player statistics, I use data from STATS LLC. Stats LLC is a leading data company
that tracks soccer players statistics across the world throughout multiple competitions including
almost every domestic soccer league, international club competitions such as the Champions
League, and international country competitions such as the World Cup. For player transfer fees, I
have already complied a list of transfer fees as reported by the British Broadcasting Company, a
reliable media outlet in Britain. As the prices are in Euros, I have used inflation data from the
European Central Bank to normalize transfer fees to a base of 2015 euros.11
As seen in Table 6, the mean age that a player is transferred is 24.48, implying that
Premier League clubs tend to purchase players before they hit their peak, usually around the age
of 27.12 More so, the average player has scored approximately 27 goals in his career before he is
transferred, but as seen, this data can vary widely from 0 goals to 165 goals. Forwards and older
players would be expected to have scored more goals in their career than defenders and younger
players. Similarly, matches started and minutes played vary widely due to differences in
experience and age.
11
European Union. Euro Inflation Over Time.
12
BBC. When do footballers hit their peak? July 2014.Newman 21
Table 6: Descriptive Statistics of Career Player Statistics for Players Transferred into the
English Premier League from 2009-201513
Variable Mean Standard Minimum Maximum
Deviation
Assists 6.75 12.83 0 116
Age 24.48 3.49 17 33
Goals 27.45 31.06 0 165
Matches Played 138.13 98.3 0 602
Matches Started 115.11 87.01 0 573
Minutes Played 10230.33 7646.73 0 50381
V. EMPIRICAL SPECIFICATION
In order to best find what factors affect a player’s transfer fee, I run regressions using the
natural log of the transfer fee as the dependent variable and player characteristics and purchasing
statistics as the independent variables, along with the intercept and error term (Formula 1).
Formula 1:
Ln(Transfer Fee) = β 0 + β 1X1+ β2X2 + ε
Where X 1 consists of player characteristics and statistics and X2 consists of purchasing
characteristics
The dependent variable in the regression is the natural log of the transfer fee. The natural
log was used in order normalize the distribution of transfer fees. The list of independent variables
13
Data from STATS LLC. Copyright 2016.Newman 22
used throughout the models can be found above in Tables 4 and 5. While one may question why
a player’s wage is excluded from this analysis, after a lot of research, there does not seem to be
any reliable website that lists wages for all Premier League players over time. Some star player’s
rumored wages are readily available, but I could not find wage data for a majority of players in
my dataset.
While it is common to analyze data through Ordinary Least Squared (OLS) regressions,
OLS estimates do not capture behavior at the tails as OLS approximates a mean of a conditional
distribution. As such, OLS estimates will not show whether or not there is a superstar effect. In
order to test for this effect, I can analyze the data using quantile regression, similar to the
analysis of Franck and Neusch. Quantile regression allows one to examine the entire distribution,
rather than solely the mean that OLS allows us to examine. This quantile regression approach
gives the possibility of testing for a superstar effect and examining the difference between the
effect of a goal on the transfer fee for a player who had a transfer fee in the 50th percentile and a
player who had a transfer fee in the 95th percentile. In my analysis below, I begin with OLS
regression and then use quantile regression to test for a superstar effect.
VI. RESULTS
a. Dependent variable of Projected Transfer Fees
To begin the analysis, the methods of Franck and Neusch’s German Bundesliga study
(2006) were used in order to extend their analysis and see how the results of the determinants of
projected transfer fees in the English Premier League compares with that of the German
Bundesliga. I began with making the dependent variable the logarithm of the projected transfer
fees of players at the time they were transferred, according to Transfermarkt. I also used many ofNewman 23
the same variables that Franck and Neusch analyzed in the regression, including goals, assists,
shots, red cards, yellow cards, age, age squared, appearances, fixed teams effects, and position
fixed effects. Of the aforementioned, Franck and Neusch found goals, assists, age, age squared,
and appearances to be significant in an Ordinary Least Squares regression with an adjusted r-
squared of 0.70. The coefficients for goals, assists, age, and appearances were al positive, while
the coefficient on age squared was negative. In their quantile regression of the 95% quantile, they
found goals, assists and age to be significant at the 1% level with a positive coefficient, and age
squared to be significant at the 1% level with a negative coefficient with an adjusted r-squared of
0.56.
Table 7: OLS Regression using Projected Transfer Fee (in Euros) as Dependent Variable14
In(ProjectedTransferFee) Coefficient SE
Age 1.252*** 0.138
Age Squared -0.023*** 0.003
Matches Played 0.002*** 0.001
Goals 0.002 0.002
Assists 0.001 0.005
Shots 0.002*** 0.001
Yellow Cards 0.002 0.002
Red Cards 0.055* 0.029
Constant -2.641 1.795
N=441
Adjusted R-Squared=0.54
Note: * signifies significance at the 10% level, ** signifies significance at the 5% level, and ***
signifies significance at the 1% level.
14
Data from STATS LLC. Copyright 2016.Newman 24
Regressing the Transfermarkt projected transfer fees against goals, assists, red cards,
yellow cards, age, age squared, and appearances against fixed team effects and position effects
led to an adjusted r-squared of 0.54 (Table 7). Age and age squared both had a p-value of 0.000,
meaning that they are statistically significant at the 1% level. Age had a positive coefficient of
1.25, which means that for each increase in age the projected transfer fee increased by 125%,
while age squared had a negative coefficient of -0.023, meaning that for each increase in age
squared the projected transfer fee decreased by 2.3%. While the increase of age on the projected
transfer fee seems high, it is important to note that the negative age squared variable is rising
much faster, mitigating much of age’s positive effect on the transfer fee. These effects happen as
there is a decreasing returns to age, due to teams wanting players who have physically developed
with more experience, but at the same time, wanting a player who is not too old given that
players tend to hit their peak in their late twenties and retire by their mid thirties (BBC). Matches
played was also statistically significant at the 1% level with a positive coefficient of 0.002,
meaning that for each match played one’s transfer fee increases by 0.2%.
Strangely, in the model, both goals and assists are incredibly insignificant with p-values
of 0.822 and 0.873, respectively. While no justification for this is apparent, especially given the
fact that Franck and Neusch found both to be statistically significant at a 1% level in their study
of the German Bundesliga, it is important to remember that these are just projected transfer fees
based on an undisclosed algorithm by Transfermarkt. More so, the Franck and Neusch study
used data from the 2004-2005 season, and perhaps Transfermarkt has changed their algorithm
since then. Also interesting was that red-cards were statistically significant at a 10% level with a
positive coefficient of 0.054, meaning that for each additional red card one receives the projected
transfer value increases by approximately 5%. This again does not make much sense, however
perhaps controlling for matches played is not enough and another measurement, such as minutesNewman 25
played, should also be controlled for, as players who play more will probably receive more red
cards.
Franck and Neusch’s study analyzes the results at the 90%, 95%, and 98% quantile to
examine if there is a superstar effect present in the transfer market. In their study, they found
that at the 90% quantile, the incremental effect of a goal leads to a 9% increase in the transfer
value, which is statistically significant at the 1% level. This is compared to a 9.5% increase at the
95% quantile and a 9.6% increase at the 98% quantile, both of which are statistically significant
at the 1% level. In my extension of this analysis to the English Premier League, goals were not
found statistically significant until the 98% quantile, where each incremental goal increased the
transfer value by 0.8%. From this, we can weakly see a super star effect with the effect of goals
projected transfer values, but it is noteworthy that goals become insignificant at the 98%
quantile. Assists were also found to be statistically significant at the 1% level with a positive
coefficient. The coefficient was 0.017 at the 90% quantile, 0.25 at the 95% quantile, and 0.029 at
the 98% quantile. In my extension to the English Premier League, assists are not statistically
significant and the coefficient has no clear pattern (Figure 6). It was also interesting to see shots
significant at the 90% quantile and the 95% quantile, but insignificant at the 98% quantile.
However, shots as a metric does not help that much, as teams would be looking for quality shots,
which are probably reflected well by the amount of goals that a player takes, over a quantity of
shots, as shots are really only beneficial to the team if they go in the net to score a goal.
The results of the English Premier League extension have the same signs that Franck and
Neusch’s analysis had and find the same variables statistically significant. However, as
previously mentioned, to truly determine how the player’s market values are affected by their
statistics, looking at the actual transfer fees is a better predictor than looking at projected transfer
values by Transfermarkt and other news outlets.Newman 26
Table 8: Chart depicting the corresponding Transfermarkt and actual market value, in Euros,
that corresponds with the respective percentile
Percentile Transfermarkt Actual Market
Value (Euros) Value (Euros)
1% 0 252,567
5% 299,250 1,011,330
10% 1,105,500 1,628,685
25% 2,645,313 2,751,000
50% 5,972,500 6,069,000
75% 11,784,250 11,612,700
90% 20,930,000 21,924,200
95% 29,6800,000 30,232,800
99% 43,310,000 42,685,470
Note: Figures adjusted for inflation to 2015 eurosNewman 27
Table 9: Quantile Regression Results using Projected Transfer Fee as Dependent Variable15
Estimation approach 90% 95% 98%
Quantile Quantile Quantile
In(ProjectedTransferFee) Coefficient SE Coefficient SE Coefficient SE
Age 0.756*** 0.196 0.698** 0.272 0.661** 0.286
Age Squared -0.015*** 0.005 -0.014*** 0.005 -0.013** 0.006
Matches Played 0.001** 0.001 0.001* 0.001 0.001 0.001
Goals -0.001 0.001 0.002 0.002 0.008*** 0.003
Assists 0.006 0.005 0.002 0.008 0.005 0.008
Shots 0.001** 0.001 0.002* 0.001 0.001 0.472
Yellow Cards 0.002 0.004 0.002 0.004 0.003 0.004
Red Cards 0.008 0.026 0.012 0.021 0.014 0.016
Constant 6.104** 0.017 7.139** 3.495 7.539** 0.040
Fixed Team Effects Yes Yes Yes
Fixed Position Effects Yes Yes Yes
N=461
Pseudo R-squared 48% 50% 54%
Note: * signifies significance at the 10% level, ** signifies significance at the 5% level, and ***
signifies significance at the 1% level.
b. Actual Transfer Fees used as Dependent Variable
To begin my analysis, I used an Ordinary Least Squared regression model with the log of
the actual transfer fees paid, in 2015 euros, as the dependent variable. In order to determine what
statistics helped explain the most variance in player’s transfer fees, I looked at player statistics
from the previous season prior to the transfer, the past three years prior to the transfer, and
15
Data from STATS LLC. Copyright 2016.Newman 28
player’s career statistics prior to their transfer. The adjusted r-squared values increased when
more years of player statistics were added. The adjusted r-squared values increased from .59
when statistics for the year prior to the transfer were used to .64 when statistics for the player’s
career prior to the transfer were used (Tables 10 through 12).
Continuing the analysis, I regressed the actual transfer fee against career statistics (Table
10). Using interaction terms, breaking goals and assists out by position, results in a similar r-
squared of 0.60 as lumping all position’s goals and assists together. In the first regression,
without the interaction terms, goals scored were statistically significant at the 1% level with a
coefficient of 0.004, meaning that each goal scored increased the player’s transfer value by 0.4%.
Forward and midfield interaction terms with goals and assists are present in the second
regression, with defense omitted to avoid multicollinearity. In the regression with the position
interaction terms, goals were only statistically significant for forwards, with each goal scored
increasing the player’s transfer value by 0.3%. As forwards are the most attacking players and
score the most goals, this result does make sense, as the role of forwards is to score goals while
midfielders and defenders have other primary responsibilities, such as creating goal-scoring
opportunities and defending.Newman 29
Table 10: OLS Regression using Statistics from Previous Season Prior to Transfer16
OLS Regression (1) (2)
In(TransferFee) Coefficient SE Coefficient SE
Goals 0.025*** 0.005
Forward Goals 0.027*** 0.008
Midfield Goals 0.016* 0.010
Assists 0.021** 0.008
Forward Assists 0.016 0.018
Midfield Assists 0.031*** 0.011
Age 0.878*** 0.122 0.865*** 0.123
Age Squared -0.017*** 0.002 -0.017*** 0.002
Minutes Played 0.000** 0.000 0.000 0.000
Fixed Purchasing Team Effects Yes Yes
Fixed Selling Team Effects Yes Yes
Fixed Window Effects Yes Yes
Constant 3.973** 1.551 4.134 1.555
N=441
Adjusted R-Squared .59 .61
Note: * signifies significance at the 10% level, ** signifies significance at the 5% level, and ***
signifies significance at the 1% level.
16
Data from STATS LLC. Copyright 2016.Newman 30
Table 11: OLS Regression using Statistics from Previous Three Seasons Prior to Transfer17
OLS Regression (1) (2)
In(TransferFee) Coefficient SE Coefficient SE
Goals 0.009*** 0.002
Forward Goals 0.011** 0.004
Midfield Goals 0.002 0.004
Defense Goals 0.016** 0.008
Assists 0.021*** 0.004
Forward Assists 0.003 0.010
Midfield Assists 0.014*** 0.004
Defense Assists -0.008 0.011
Age 0.802*** 0.119 0.771*** 0.010
Age Squared -0.016*** 0.002 -0.016*** 0.002
Ln(Minutes Played) 0.035 0.440 0.016 0.003
Fixed Purchasing Team Effects Yes Yes
Fixed Window Effects Yes Yes
Fixed Selling Team Effects Yes Yes
Constant 4.493*** 0.094 4.809*** 1.559
N=441
Adjusted R-Squared .61 .61
Note: * signifies significance at the 10% level, ** signifies significance at the 5% level, and ***
signifies significance at the 1% level.
17
Data from STATS LLC. Copyright 2016.Newman 31
Table 12: OLS Regression using Career Statistics Prior to Transfer18
OLS Regression (1) (2)
In(TransferFee) Coefficient SE Coefficient SE
Goals 0.004*** 0.001
Forward Goals 0.003** 0.098
Midfield Goals 0.003 0.226
Assists 0.010*** 0.003
Forward Assists 0.018** 0.010
Midfield Assists 0.008*** 0.004
Age 0.742*** 0.119 0.771*** 0.010
Age Squared -0.016*** 0.002 -0.016*** 0.002
Ln(Minutes Played) 0.130*** 0.440 0.016 0.003
Fixed Purchasing Team Effects Yes Yes
Fixed Selling Team Effects Yes Yes
Fixed Window Effects Yes Yes
Constant 4.468*** 1.456 4.692*** 1.459
N=441
Adjusted R-Squared .64 .64
Note: * signifies significance at the 10% level, ** signifies significance at the 5% level, and ***
signifies significance at the 1% level.
Career assists are also statistically significant in the regression. For each assist a player
has, his transfer value increases by 1%. Adding in interaction terms to the regression, we see that
assists are statistically significant for forwards and midfielders. This again makes sense, as
forwards and midfielders are primarily responsible for creating goal-scoring opportunities. An
18
Data from STATS LLC. Copyright 2016.Newman 32
incremental assist by a forward leads to a 1.8% increase in the transfer value, while an
incremental assist by a midfielder leads to a 1% increase in the transfer value. It is interesting to
see that the coefficient on the assist variable is greater than that on the goal variable. Perhaps
English Premier League clubs value play-making ability and players who create goal-scoring
opportunities and the assist variable is a good proxy for this.
Both “Age” and “Age squared” are statistically significant with 0.000 p- values in both
regressions. In the non-interaction term regression, the coefficient on “Age” illustrates that, for
every yearly increase in a player’s age, the transfer fee increases by 74.2 percent. However, the
coefficient on “Age squared” is -0.016, meaning that there is a 1.6 percent decrease in the
transfer value for every increase in the “age squared” variable. In the model with interaction
terms, an increase in the player’s age leads to an increase in the transfer fee by 77.1% and an
incremental increase in the age squared variable leads to a 1.6% decrease in transfer value.
However, a one-unit increase in the age-squared variable will not occur. For example, an
increase in age from 25 to 26 will lead to a 51-unit increase in the age-squared variable, leading
to an 81.6 percent decrease in transfer value. When combined with the 77.1% incremental
increase in transfer value per year, from the interaction term model, the net effect for the age
variables sees a 4.5% decrease in transfer value.
While seemingly counterintuitive, it is reasonable to believe that there are diminishing
returns to age. Although age can also be considered a proxy for experience, which can be a good
thing, as a team would like someone in the prime of their career, there is a drawback to the
purchase of older players. At the same time, teams do not want to pay for someone who is too
old to play at a highly competitive level. For example, I analyzed the effect that age has on a
player using the OLS regression model in Figure 2. I analyzed the variables at their mean, and
used a forward who was purchased by Arsenal for the team and position dummies. The figureNewman 33
illustrates the effect, and shows a peak of value at approximately 24 years old. After this age,
there is a negative effect on the transfer fee. A BBC study showed that players tend to reach
their peak form at approximately 27 to 29 years old.19 A purchasing club would want to
maximize the amount of time that they have their players in peak form, so the maximum transfer
value occurring around the age of 24 years old, all else equal, seems reasonable.
The natural log of career minutes played variable was interestingly significant in the non-
interaction term model, but was not significant when interaction terms were included. A one-
percent change in the natural log of minutes played is associated with a 0.13% increase in a
player’s transfer value. This variable serves as a proxy for experience and talent, as a player who
has played frequently with their previous teams would most likely be talented, or else they would
have not been selected to play as much as they were.
Figure 2: Illustrating the Effect of Age on Transfer Fee
How
Transfer
Fee
Varies
By
Age
25.8
25.7
Ln(Transfer
Fee)
25.6
25.5
25.4
25.3
25.2
ln(TransferFee)
25.1
25
24.9
18
19
20
21
22
23
24
25
26
27
28
29
30
Age
19
BBC. When Do Footballers Hit Their Peak? July 2014.Newman 34
Table 13: Quantile Regression of Transfer Fees with Career Statistics20
Estimation 50% 75% 90% 95%
approach Quantile Quantile Quantile Quantile
In(TransferFee) Coefficie SE Coefficient SE Coefficient SE Coefficient SE
nt
Age 0.771*** 0.193 0.765*** 0.181 0.684*** 0.162 0.696*** 0.174
Age Squared -0.016*** 0.004 -0.016*** 0.004 -0.015*** 0.003 -0.015*** 0.004
Assists (Forwards) 0.014 0.001 0.014 0.021 0.004 0.019 0.006 0.003
Assists (Midfield) 0.008* 0.005 0.007* 0.004 0.006 0.004 0.007* 0.003
Goals (Forwards) 0.003 0.002 0.005** 0.002 0.006** 0.003 0.007** 0.003
Goals (Midfield) 0.001 0.001 0.004 0.003 0.002 0.002 0.005 0.003
ln(Career Minutes) 0.110*** 0.027 0.095*** 0.037 0.045 0.045 0.000 0.040
Constant 4.817** 2.453 5.50*** 2.267 7.036*** 2.001 7.849*** 2.714
Purchasing Effects Yes Yes Yes Yes
Selling Effects Yes Yes Yes Yes
Window Effects Yes Yes Yes Yes
N=441
Pseudo R-squared 47% 50% 54% 55%
Note: * signifies significance at the 10% level, ** signifies significance at the 5% level, and ***
signifies significance at the 1% level.
20
Data from STATS LLC. Copyright 2016.Newman 35
c. Quantile regression
Quantile regression estimates a conditional quantile of a response variable given
covariates using bootstrap replications. This allows one to examine the incremental impact of
different statistics, such as goals, at different percentiles to determine if there is a superstar effect
present in the English Premier League transfer market and if so, what that impact is. In the
model, we primarily see the superstar effect present with regards to forward goals, which is
similar to the results that Franck and Neusch found in their 2006 study. In the model, we see that
for a forward at the 50% quantile, an incremental goal increases their transfer value by 0.3%.
However, for a forward at the 95% quantile, an incremental goal increases his transfer value by
0.7%. This effect increases throughout the model, illustrating the superstar effect for goals scored
by forwards. It is noteworthy that this magnitude of this effect is smaller than the magnitude that
Franck and Neusch saw in the German Bundesliga of approximately nine percent; however, it is
once again important to note that they used projected transfer fees instead of actual fees.
However, other statistics do not exhibit a clear superstar effect in the model. Franck and Neusch
did not separate goals and assists out by position, but found that total goals was the only statistic
that exhibited a Superstar effect in the German Bundesliga.
The downward trend on the coefficient of the natural log of minutes is also noteworthy.
From the regression, it appears that there is a backwards superstar effect, where having played
less minutes increases the transfer fee by the most in the lower quantiles. However, the variable
is not statistically significant at the 90% and 95% levels, making this a harder trend to read. On a
similar note, the coefficients on age decrease from the 50% quantile to the 75% quantile to the
90% quantile, but slightly increase from the 90% quantile to the 95% quantile. I plan on
examining these trends closer in the coming weeks to see if anything can be extracted.Newman 36
It is also interesting to note that the psudeo r-squared increases throughout the model. As
previously noted, Transfermarkt poorly predicts larger transfer fees; however, an increase in the
psudeo r-squared throughout the model illustrates that the above model in Figure 11 explains
more of the variation among larger transfer fees than it does with smaller transfer fees.
d. Addition of Homegrown Variables in Model
When added to the regression, homegrown variables have p-values that are not
statistically significant at a 10% level. Neither a homegrown dummy nor homegrown position
interaction dummies are significant at any reasonable level, which is surprising given the
artificial eight person homegrown quota that all clubs must fulfill. In regression one, seen in
Table 12, where a homegrown dummy is added to the regression, the p-value is 0.167, making it
an insignificant variable. Similarly, when interaction terms are formed, between homegrown
status and position, the variables are also not significant at a 10% level. The p-value for the
homegrown forward variable is 0.263 while the p-value for the homegrown midfield variable is
0.011.Newman 37
Table 14: OLS Regression using Career Statistics Prior to Transfer and Homegrown Variables21
OLS Regression (1) (2)
In(TransferFee) Coefficient SE Coefficient SE
Goals 0.006*** 0.002 0.007*** 0.0013
Assists 0.006*** 0.003 0.007*** 0.003
Age 0.701*** 0.105 0.661*** 0.107
Age Squared -0.015*** 0.002 -0.014*** 0.002
Ln(Minutes Played) 0.126*** 0.021 0.135 0.021
Fixed Purchasing Team Effects Yes Yes
Fixed Selling Team Effects Yes Yes
Fixed Window Effects Yes Yes
Constant 4.350*** 1.490 4.917*** 1.512
English Homegrown 0.105 0.076
Homegrown Forwards 0.162 0.145
Homegrown Midfielders 0.181 0.121
N=441
Adjusted R-Squared .64 .64
Note: * signifies significance at the 10% level, ** signifies significance at the 5% level, and ***
signifies significance at the 1% level.
21
Data from STATS LLC. Copyright 2016.Newman 38
e. Addition of Competition Level Variables to Model
Another variable that was hypothesized to be statistically significant was the competition
level. In order to analyze competition level, I used two separate variables, an English Premier
League dummy if the player was transferred from a Premier League club as well as a Top Five
League Dummy variable if a player was transferred from a top five league in the world, based on
historical UEFA league coefficients.22 The top five league variable includes the English Premier
League, the German Bundesliga, the French Ligue 1, the Italian Serie A, and Spain’s La Liga. As
seen in Table 13, the p-value for the Premier League sale is .333 in regression one while the p-
value for the top five sale is 0.350 in regression two, making both variables insignificant at the
10% level. One potential reason that these competition level variables are statistically
insignificant is that due to increased international competitions, such as the Champions and
Europa Leagues at the club level, and continental and world tournaments such as the European
Championship and World Cup at the country level, scouts are able to see players perform against
quality sides. More so, there is a subjective quality to scouting that statistics do not fully track,
such as ball control, which is why scouts are constantly sent to games to scout players instead of
solely relying on their statistics.23 Another explanation could be that talented young players tend
to move to somewhat less competitive leagues where they can have a significant amount of
playing time rather than sitting on the bench for a team in a top league. Examples of this can be
seen by Gedion Zelalem’s loan from Arsenal of the English Premier League to Rangers of the
Scottish Championship, Michael Bradley’s transfer from Roma of Serie A to Toronto FC of
22
UEFA. UEFA Rankings.
23
LA Times. Europeans Continue to Tap US Soccer Pool. April 2012.
Harvard Sports Analysis. Team Form Recency Bias and Regression to the Mean. August 2015.Newman 39
MLS, and Adnan Januzaj’s loan from Manchester United of the English Premier League to
Borussia Dortmund of the German Bundesliga.
Table 15: OLS Regression using Career Statistics Prior to Transfer with Competition Level
Variables24
OLS Regression (1) (2)
In(TransferFee) Coefficient SE Coefficient SE
Goals 0.006*** 0.001 0.006*** 0.001
Assists 0.006*** 0.003 0.006** 0.003
Age 0.707*** 0.106 0.684*** 0.186
Age Squared -0.015*** 0.002 -0.015*** 0.002
Ln(Minutes Played) 0.122*** 0.022 0.132*** 0.022
Constant 4.306*** 1.492 4.532*** 1.052
Competition Level
Top Five League 0.071 0.076
England 0.072 0.072
Fixed Purchasing Team Effects Yes Yes
Fixed Selling Team Effects Yes Yes
Fixed Purchasing Window Effects Yes Yes
N=441
Adjusted R-Squared .64 .64
24
Data from STATS LLC. Copyright 2016.Newman 40
f. Addition of Form Variables to Model
In order to test whether or not a player’s form was significant prior to the transfer, two
sets of form variables are tested in the regression. In the first regression, assist and goal
difference variables are included interacted with a player’s position. These variables represent
the net change in goals and assists in the season prior to the transfer compared to the season two
years prior to the transfer. These can be positive or negative, depending on how the player’s
performance was compared to two years before he was transferred. These variables are all
statistically insignificant at a 10% level. In regression two of Figure 14, simple dummy variables
are added to the regression that are equal to one if a player increased his output in forwards or
assists and equal to zero if that was not the case. Again, these variables are insignificant at the
10% level. While one would expect form to play a component in a player’s transfer fee, it is
possible that teams take a more holistic approach to the transfer to insure that a player’s run of
form is not a fluke. More so, players and teams typically regress to a mean, as shown by Harvard
Sports Analysis, so teams may not want to base a decision off of just eight months of play.25
25
Harvard Sports Analysis. Team Form Recency Bias and Regression to the Mean. August 2015.Newman 41
Table 16: OLS Regression using Career Statistics and Form Variables26
OLS Regression (1) (2)
In(TransferFee) Coefficient SE Coefficient SE
Goals 0.006*** 0.001 0.006*** 0.001
Assists 0.007** 0.003 0.006* 0.003
Age 0.695*** 0.106 0.699*** 0.106
Age Squared -0.015*** 0.002 -0.015*** 0.002
Ln(Minutes Played) 0.126*** 0.021 0.126*** 0.021
Constant 4.439*** 1.496 4.351*** 1.501
Form
Assist Change (Forward) 0.012 0.014
Assist Change (Midfield) 0.004 0.011
Goal Change (Forward) 0.059 0.009
Goal Change (Midfield) 0.008 0.009
Assist Increase Dummy (Forward) 0.018 0.124
Assist Increase Dummy (Midfield) 0.078 0.095
Goal Increase Dummy (Forward) 0.058 0.125
Goal Increase Dummy (Midfield) 0.024 0.0944
Fixed Purchasing Team Effects Yes Yes
Fixed Selling Team Effects Yes Yes
Fixed Purchasing Window Effects Yes Yes
N=441
Adjusted R-Squared .64 .64
26
Data from STATS LLC. Copyright 2016.Newman 42
VII. OVERALL CONTRIBUTION AND FURTHER EFFECTS
The regression model built gives fans, players, clubs, and the media a better picture as to
what certain players are worth in today’s market in the English Premier League, a study that has
not been publicly completed before. In the base OLS regression model we see that goals, assists,
age, and minutes played have a positive coefficient and are statistically significant, while age-
squared has a negative coefficient due to the fact that players hit a performance peak in their late
twenties. We see that a superstar effect is evident in the soccer market, as goals scored by
forwards are worth more to a player’s transfer value when they are scored by elite players, rather
than average players. While the study explains much of the variation in transfer fees of English
Premier League players, it is missing a popularity factor to compensate for the popularity of a
player. As the sport has become more global over the past ten years, and can now be easily seen
on television in countries such as Chia, India, and the United States, I would imagine that
popularity matters more than ever to give teams a boost in merchandise sales and television
ratings. With more time, analyzing the popularity of a player should be beneficial in determining
transfer values of players, but the model at the current state illustrates the impact of on-field
performance on the transfer value of outfield players.You can also read
